運用AI BERT探究臺灣新聞媒體之框架：以2020年美國總統選舉議題為例

Using AI BERT to Investigate the Framework of Taiwan News Media: A Case Study of the 2020 United States Presidential Election

美國在全球事務向來具有舉足輕重之影響力，在臺灣與中國的關係間更扮演重要角色，而近年來美中抗爭加劇，也促使美國對於臺灣之態度與政策更顯得重要。因此，本研究透過分析《TVBS新聞網》、《三立新聞網》、《ETtoday新聞雲》、《東森新聞》、《蘋果新聞網》、《聯合新聞網》與《自由時報電子報》七家網路新聞媒體，透過大數據分析取徑對文本進行探究，並藉由Google BERT模型將文本分類為「賽局框架」、「策略框架」、「問題框架」、「中臺關係框架」和「與選舉無關」五個類目，同時針對人工編碼與Google BERT模型之分類準確率差異進行比較，來探討過去媒體研究指出不同新聞媒體間，經常會受到媒體利益或立場影響，進而產生出具差異性的論述。最終研究結果顯示，本研究透過人工標記資料微調後產出之BERT模型，與人工標記之結果類似，準確率達84.48%，並成功預測出三家新聞媒體的排序。而在七家臺灣主流網路媒體，對於2020年美國總統選舉報導的呈現，透過新聞框架之比較，本研究發現媒體立場確實會造成報導內容產生差異的現象。

Background: The Taiwanese people are generally very concerned about U.S. presidential elections, but their views on them are different from most countries’social phenomena. Past research has pointed out that the media’s discussion of the issues will affect the audience. Influence and other factors have prompted this study to take Taiwan news media as the analysis object, in view of the strategic and economic importance of the United States in regards to the relationship between the two sides of the Taiwan Strait. How Taiwanese media present the issues of U.S. presidential elections are certainly worth exploring. Purpose: With the advancement of technology, more and more experiments expect to use technology to assist in research, and many researchers have begun to use computer programs to analyze data based on the advantages of big data. This study believes that the approach of big data research can help scholars sort out the topical framework of news media and understand the core meaning behind it. Therefore, this study compares the differences between the Google AI BERT model and manual classification, so as to confirm whether the error value between the latter and the former affects the research results, and proposes the verification results of the application of the classification model in the field of data analysis. RQ1: Using the Google AI BERT classification model for the 2020 U.S. presidential election report, what is the accuracy of the classified subject frame? RQ2: What thematic framework is used to reproduce the 2020 U.S. presidential election among different online news media? RQ3: What is the significance of the differences in the thematic framework of coverage of the 2020 U.S. presidential election between different online news media? Method: This study uses big data analysis to analyze texts and applies Python to scrutinize news frames to understand the meaning behind a large amount of information. In the end, 35,991 news samples in total were collected from seven online media firms. Reports of the U.S. presidential election have established five mutually exclusive categories: Game Framework, Strategy Framework, Issue Framework, China-Taiwan Relationship Framework, and Unrelated to Election. In this study, 3400 items were extracted from each of the five categories that were manually classified as the training dataset of the AI BERT model. The parameters were set as Batch size=16, learning rate=1e-5, and epoch=6, and the accuracy of the model training results was 84.48%. Thus, the study adopted this model as the classification model for subsequent studies. Findings: From the artificially encoded data in Figure 4 to Figure 10, it can be observed that only ETTV News has a relatively average reporting ratio among the news media, and the media’s“Game Framework”ratio is also the highest among the seven. In addition, in the“China-Taiwan Relationship Framework”, SET News, ETtoday News, Apple Daily, and LTN News are all over 30%, while SET News’number of news related to the“China-Taiwan Relationship Framework”is even higher than 43%, or greater than that of other news media. This means that when SET News reports on the U.S. presidential election, it particularly focuses on content related to China or Taiwan. TVBS News, ETTV News, and UDN News accounted for nearly 25%, of which TVBS News accounted for only 22%. Among them, UDN News obviously focuses on“Issue Framework”, which accounts for 43%. In“Game Framework”media, SET News, Apple Daily, ETtoday News, UDN News, and LTN News accounted for only less than 10%.